>>66451435>>66450733This is without LoRA. AnimagineV3 has been trained on 1.2 million images on Danbooru, so most of the bigger vtubers should have been ingested by the model.
Note that this model expects a certain order of tags (the same as in NAIv3), but in return the output seems to be more accurate (
https://huggingface.co/cagliostrolab/animagine-xl-3.0#tag-ordering ), see:
https://litter.catbox.moe/2psmzx.webphttps://litter.catbox.moe/eiwbqr.webpThis one with tag ordering: 1girl, ninomae ina'nis, final fantasy xiv, fate/grand order, [rest of the prompt], best quality, masterpiece
I found that the "from what series" tag can have an additional effect on the style, e.g. using Oshi no Ko as series indicator makes it easier to get this specific eye twinkle effect and down the line improves eye detail.